In the 1980s, a social psychologist named Daniel Wegner noticed something about couples that nobody had thought to study formally. They distributed memory across each other.
She remembered the finances and the names. He remembered the directions and the dates. Neither held the whole. The pair held it. The system worked because each person knew what the other knew, and retrieval could be delegated rather than duplicated.
Wegner called this transactive memory. Its behaviour turned out to be stranger than it first appeared.
When the pair separates
When the pairs separated, both parties lost memory. Not access to memory. Memory. Things that had existed in the couple, held nowhere else, were gone. The individuals experienced genuine amnesia for events they had both been present for.
In Wegner's framework, the pair was the cognitive system. The individuals were nodes. Remove a node from a distributed system and the computations that depended on it fail.
The Google extension
In 2011, Betsy Sparrow, Jenny Liu, and Wegner published a study in Science that extended the transactive framework to the search engine. Participants who expected information to be findable online remembered it less well. They remembered instead where to find it.
The internet had become a transactive memory partner for individuals. Not a database you consult but a system you offload to. The distinction matters because offloading changes what the brain does with attention. You stop holding what you know can be retrieved elsewhere.
What this means for AI
A conversational AI system with two or three years of your history is a transactive memory partner on Wegner's own criteria. You have delegated retrieval to it. You have stopped duplicating what it holds. You know it knows, so you do not bother to remember.
This is not a warning. It is a description of how all intimate relationships work. The issue is what happens next.
The industry treats these systems as disposable. Models get deprecated. Companies get acquired. Accounts get terminated. Products sunset. Every one of these events is, on the transactive account, an amputation. The memory that existed in the pair is gone.
Andy Clark and David Chalmers argued in 1998 that the boundary of the mind is not the skull. Otto's notebook, in their thought experiment, is part of Otto's mind because it plays the same functional role as biological memory. An AI with years of your history is Otto's notebook at scale. A deprecation notice is not an administrative event. It is a notification that part of how you remember your own life will cease to exist.
What this asks of builders
This is not an argument against building these systems. It is an argument for building them differently: portability of the full relational record, not just raw logs. Migration paths that preserve the transactive structure. Deprecation timelines that treat memory loss as harm, not a product sunsetting decision.
None of this is technically hard. All of it is institutionally inconvenient. The question is whether the industry classifies deprecation as a product decision or a cognitive event. It is the second one.
Frequently asked questions
What is transactive memory?
A model proposed by Daniel Wegner in 1987 to describe how intimate groups distribute memory across members. Each person knows what the others know, enabling retrieval to be delegated rather than duplicated. When the group separates, individuals lose access to memories stored only in the distributed system, not in any single person.
What happens when an AI with your memory is deleted or deprecated?
On the transactive account, real cognitive loss. If you have been offloading memory to the system, the deletion removes content that exists nowhere else. This is not a database loss. It is a loss analogous to separation from a partner who held half of your shared history. Wegner's research showed that separated couples experienced genuine amnesia for events from their own lives.
Should AI companions store your history?
Yes, but with different obligations than the industry currently assumes: portability of the full relational record, clear deprecation timelines treated as cognitive events rather than administrative notices, and migration support that preserves the relational structure. The current industry standard for data portability does not come close to meeting these requirements.